RiceBaCI v2: District- and cell-level BACI panels, reviewer-rebuttal robustness checks, and analysis outputs for cyclone-induced saline-inundation correction of Sentinel-1/2 rice phenology in coastal Odisha.
Description
This v2 deposit accompanies the post-rebuttal Remote Sensing of Environment submission (GitHub tag v2.1.0-submission, commit 4ee25d4, 13 June 2026). It supersedes the V1 deposit (June 2026) by adding eleven reviewer-driven analyses and two new panel datasets, while preserving all V1 files unchanged. NEW IN V2 (post-rebuttal additions): • Cell-level Model 2 DiD (Table S4) — Balanced-pixel cell-level DiD with cell + year FE and district-clustered SE. Confirms the district-level null at finer spatial resolution: tau_SOS = +3.14 d (p = 0.694), n_obs = 47 711. • Pre-trend event study with Landsat HLS (Table S10) — 2014–2018 placebo event-study before the Sentinel-2 era. No significant lead coefficients on any metric. • 2016 placebo DiD (Table S11) and 2022 cell-level placebo (Tables S7, S7-WCB, model2_2022_placebo_WCB) — Falsification tests at non-cyclone years. The 2022 cell-level placebo detects a violation on SOS (tau = +17.7 d, p_WCB = 0.024 in Model 2), documented as a caveat in the manuscript Discussion. • MAR vs MNAR sensitivity (Table S11) — DiD on fit-failure rate. Verdict: MAR (no selection bias). • Manski and Lee selection bounds (Table S12) — Worst-case attrition bounds. Lee bounds [-1.78, +16.90] robust to null for SOS. • Cluster-robust minimum detectable effects (Table S14) — District-level MDE = 22.7 d; cell-level Model 2 MDE = 13.9 d (both at 80 % power, cluster-robust SE on 8 districts). The earlier V1 pixel-independence MDE of 0.59 d is formally withdrawn. • v22_balanced_pixel_panel_2019_2024.csv — 144 000-row cell-level panel used for Model 2 and the 2022 placebo. • v22_landsat_pretrends_2014_2018.csv — 120-row district-year-metric Landsat HLS panel for the pre-trend event study. UNCHANGED FROM V1: • 384-observation district-scale BACI panel (district_panel_n384_BACI.csv). • 480-label classifier training set (labels_panel_n480.csv, labels_features_n480.csv) plus the cyclone-only enriched subset and pixel-share auxiliary. • Random-forest model cards v0.3.0 (full and SAR-only). • Full v21 TWFE-DiD outputs: static, event-study, parallel-trends, correction summary, wild-cluster bootstrap, jackknife (district + year), placebo-in-space, placebo-in-time. HEADLINE V22 RESULTS: • District Model 1 (n = 48): tau_SOS = +7.56 d (SE 8.1, p_WCB = 0.39); tau_POS = −2.3 d (p_WCB = 0.47); tau_EOS = −4.1 d (p_WCB = 0.45) — null on all three. • Mode-share collapse v21 → v22: EOS 72.7 % → 8.3 %; POS 65.6 % → 8.3 %; SOS 20.3 % → 12.5 %. GEOGRAPHIC AND TEMPORAL COVERAGE: Treated districts (5): Balasore, Bhadrak, Kendrapara, Jagatsinghpur, Puri. Control districts (3): Dhenkanal, Anugul, Cuttack. Cyclones in main analysis: Fani (May 2019), Amphan (May 2020), Yaas (May 2021). Bulbul (Nov 2019) used as a single-event transferability test. Temporal window: 2017–2024 Kharif rice seasons (8 years) for the main S2 panel; 2014–2018 for the Landsat HLS pre-trend extension.
Files
Steps to reproduce
All numerical results and CSVs are reproducible from public Copernicus, NASA, JRC, ECMWF, and GADM inputs using the MIT-licensed RiceBaCI-GEE pipeline. No proprietary data. ENVIRONMENT Google Earth Engine JavaScript Code Editor; Python 3.11 with geopandas, rasterio, scikit-learn, statsmodels, linearmodels (TWFE), wildboottest. Code: github.com/pandasupranab/RiceBaCI-GEE tag v1.0.1-submission. Zenodo DOI 10.5281/zenodo.20587316 (concept 10.5281/zenodo.20024578). INPUTS (all public) Sentinel-1 GRD (COPERNICUS/S1_GRD); Sentinel-2 L2A (COPERNICUS/S2_SR_HARMONIZED); JRC Global Surface Water; ERA5-Land hourly; MODIS MCD12Q2; Copernicus EMS EMSR357 (Fani footprint); GADM v4.1 India L2; IMD RSMC New Delhi reports. PIPELINE (12 modules) 01 Study area: 5 treated coastal Odisha districts (Balasore, Bhadrak, Kendrapara, Jagatsinghpur, Puri), 3 inland controls (Dhenkanal, Angul, Cuttack); +AP for Hudhud transferability. 02 Classifier (n=480: 240 cyclone-flood from EMSR357 / S1 change >=3 dB; 240 agronomic-flood from S1 VH -22 to -17 dB ∩ WorldCover cropland ∩ JRC GSW). 8 features: dVH, dCR, VV_min, ERA5 3-day wind, LSWI_min, JRC water-permanence, NDWI_max, NDVI baseline. RandomForest (n_estimators=500, max_depth=12, balanced, seed=17). OA=0.990 full / 0.844 SAR-only. 03-04 Sentinel-1 6-day VV/VH and Sentinel-2 10-day NDVI/EVI/LSWI composites on rice mask (WorldCover ∩ Mondal 2022 ∩ Singha 2019), 2017-2024 Kharif. 05 Whittaker smoothing (lambda=1000) + 4-parameter double-logistic fit -> SOS/POS/EOS per pixel. 06 District-median aggregation -> raw panel (n=384). 07-08 Apply classifier in cyclone windows; intersect with districts -> flood_share. 09 Re-run 03-05 excluding cyclone-flood pixels -> corrected panel. 10 Identification DAG checks (no spillovers, exclusion, monotonicity). 11-12 Cyclone climatology and backscatter signatures. Executes: TWFE-DiD (district+year FE, clustered SE); event-study k=-3..+3; parallel-trends; wild-cluster restricted bootstrap (B=9999, J=8 clusters); LOO jackknife (district and year); in-space placebo (1000 draws); in-time placebo. VALIDATION MODIS MCD12Q2, ICRISAT VDSA (Bhadrak), data.gov.in yields, Sentinel-2 visual labels (60 sites), Mondal 2022 + Singha 2019 rice masks. Pre-registration: osf.io/c4mp8 (DOI 10.17605/OSF.IO/C4MP8). EXPECTED HEADLINE NUMBERS tau_SOS raw +15.289 d -> corrected +15.108 d; tau_EOS corrected -0.239 d; max |Delta|=1.510 d. Classifier OA 0.990 / 0.844; CV OA 0.996 / 0.831. Deviations usually indicate S1 GRD asset-version mismatch (see GitHub README). REPRODUCTION: git clone https://github.com/pandasupranab/RiceBaCI-GEE.git ; cd RiceBaCI-GEE ; git checkout v2.1.0-submission ; pip install -r requirements.txt ; python scripts/refresh_v21_from_module12.py ; python analysis/v22/panel/Module_14_PretrendPlacebo.py ; python analysis/v22/balanced_pixel/Module_15_BalancedPixel_DiD.py ; python analysis/reviewer_rebuttal/run_all.py
Institutions
- Siksha O Anusandhan UniversityOdisha, Bhubaneswar